Inversion of Toeplitz Matrices With Only Two Standard Equations George Labahn* Department of Computer Science University of Waterloo Waterloo, Ontario, Canada N2L 3Gl and Tamir Shalom Department of Computer Science Columbia University New York, New York 10027 Submitted by Hans Schneider ABSTRACT it is shown that the invertibility of a Toeplitz matrix can be determined through the solvability of two standard equations. The inverse matrix is represented by two of its columns (which are the solutions of the two standard equations) and the entries of the original Toeplitz matrix. 1. INTRODUCTION Let A be an n-by-n Toeplitz matrix: a-1 aP2 *‘a a-(,- 1) a0 a_l *** a-(“-?_) A= a1 a, ... a-(“-,) f ! .‘. ! npl an-2 an-3 ... a0 *Partially supported by NSERC operating grant 6194 LINEAR ALGEBRA AND ITS APPLICATIONS 175: 143-158 (1992) 0 Elsevier Science Publishing Co., Inc., 1992 143 655 Avenue of the Americas, New York, NY 10010 0024-3795/92/$5.00
Transcript
Inversion of Toeplitz Matrices With Only Two Standard Equations
George Labahn*
Department of Computer Science
University of Waterloo
Waterloo, Ontario, Canada N2L 3Gl
and
Tamir Shalom
Department of Computer Science
Columbia University
New York, New York 10027
Submitted by Hans Schneider
ABSTRACT
it is shown that the invertibility of a Toeplitz matrix can be determined through
the solvability of two standard equations. The inverse matrix is represented by two of
its columns (which are the solutions of the two standard equations) and the entries of the original Toeplitz matrix.
1. INTRODUCTION
Let A be an n-by-n Toeplitz matrix:
a-1 aP2 *‘a a-(,- 1)
a0 a_l *** a-(“-?_)
A= a1 a, ... a-(“-,)
f ! .‘. ! npl an-2 an-3 ... a0
*Partially supported by NSERC operating grant 6194
LINEAR ALGEBRA AND ITS APPLICATIONS 175: 143-158 (1992)
0 Elsevier Science Publishing Co., Inc., 1992
143
655 Avenue of the Americas, New York, NY 10010 0024-3795/92/$5.00
Xl
II!_ 0 0 Yn Yn-1 **. Yl
A-&-L x2 Xl . . . 0 0 Yn ... Yz
Xl
xtl x n-l Xl
0 . . . 0
Yl . . . 0
-
.
.Yn-1 Yl
Another case in which two columns of the inverse of a regular Toeplitz matrix are sometimes sufficient to represent the entire inverse matrix appears in [2].
144 GEORGE LABAHN AND TAMIR SHALOM
where a._(._,), . . . , a,_ 1 are complex numbers. We use the shorthand A = (a, _ >;, 4 = 1 for a Toeplitz matrix.
TK e results of Gohberg and Semencul [3] show that the inverse of a regular Toeplitz matrix can be sometimes represented via its first and last columns. Their theorem is as follows:
THEOREM 1.1 (Gohberg and Semencul). Let A = Cap_,>,” 4= 1 be a Toeplitz matrix. If each of the systems of equations
(p=L% . . . . ft)
is solvable and xl # 0, then A is invertible and
THEOREM 1.2 (Gohberg and Krupnik). Let A = (a,_ y>p”, y = I be a Toeplitz matrix. lf each of the systems of equations
(p = l,%...,n),
INVERSION OF TOEPLITZ MATRICES
is solvable and x, z 0, then A is invertible and
. . .
X1X" X1X,-l
x2 *n XZXn-1
+ . [l ( xrl XII X,X,-l
0 0 2” 0 : : II , 0 0
*1 0 0
.., x1x1
. . . x2*1
“. : I;
. . . X,X1
. . . 22
:j
2,
0
145
(14
However, this is not always the case. It is shown in [2] that the inverse of the Toeplitz matrix
cannot be determined by any pair of its columns. Ben-Artzi and Shalom have shown in [l] that three columns of the inverse
of a regular Toeplitz matrix are always enough to reconstruct it.
THEOREM 1.3 (Ben-Artzi and Shalom). Let A = (a,_ 4);. 4= 1 be a Toeplitz matrix. lf each of the systems of equations
(p = l,%...,n), q=l
q=l
q=l
146 GEORGE LABAHN AND TAMIR SHALOM
is solvable for any integer number 1 (1 < 1 < n) for which x1 # 0, then A is
invertible with inverse given by
+
21 . . . 0 0
:I[
0 x, ... x2
Z2-yi e.0 0 0 : : . . :
t (j 0 ..: x.”
Zll-YYn-1 *** zz-Yl Zl 0 0 *** 0 11 . (1.2)
The results of [8] show that the inverse of a Toeplitz matrix A can always be determined by only two systems of equations having A as coefficient matrix. Subsequent results of [4], [7], and [6] show that one of the two equations can be chosen to be of the form
AX = E, (1.3)
where E is a column vector in the standard basis in C”. In all four cases, however, the second equation has a right-hand side that depends on the entries of the matrix A.
The main result of this paper shows that it is enough to solve only two systems of linear equations of the form (1.3) where E is a column vector in the standard basis in @” in order to determine the invertibility of A. In this case, we obtain a formula for A - ’ in terms of two of its columns and the entries of A.
THEOREM 1.4. The Toeplitz matrix A = (a, _ ,I,“, y = 1 is invertible if and only if the following conditions hold:
(a> There exists a solution for the following system of equations:
k aP-4xy = $,, (p = 1,2 ,..., n). q=l
(b) Let 1 be such that x, z 0 and xy = 0 for all q > 1. Then there exists a solution for the following system of equations:
2 ap-vzy = $?.“f2H (p = 1,2 ,..., n). 9=1
given by (1.2) for
Yl Xl YZ II : [:i = s - “.” [U-I a_2 0.. a-(,-l) 0
I
Y” x,,
INVERSION OF TOEPLITZ MATRICES 147
In case 1 = n then A-’ is given by (l.l), while in case 1 < n then A-’ is
n-l
I
Here S denotes the n-by-n lower shifi matrix
In the next section we prove Theorem 1.4 and obtain some corollaries. In Section 3 we state the analogous results for Hankel matrices, and in Section 4 we consider the issue of minimal&y. Namely, when can the solution of a unique system of linear equations AX = E determine the invertibility of the Toeplitz (or Hankel) matrix A?
2. PROOF OF THE MAIN RESULT
First let us present some useful notation. We denote the row with entries
b,, b,, . . , b,Y either by row(b,, b,, . . , b,%) or by row(b,Y,,. The column with entries b,, b,, . . , b,T is denoted either by col(b,, t)s,. . , b,) or by col(b,)f= 1. Let E(l), EC’), . . , E’“’ and F(l), F(s), . . , F(“) be the n unit columns and unit rows, i.e.,
and
Ecy) = c01($,<,);=~ (4 = 1,2,...,n)
Fey) = row(S,,y)C=, (4 = 1,2 ,..., n).
We extend the definition in the sense that E(O), for example, is the zero column.
For any matrix A we denote its transposed matrix by AT.
148 GEORGE LABAHN AND TAMIR SHALOM
LEMMA 2.1. Let A = (u~_~);,~=~ be a Toeplitz matrix and X = col( xp>p”= 1 be a column vector for which
Define the column vectors
M zr EC’), (2.1)
Xci) = (S - XF(‘)AS)iX (i > 0). (2.2)
Note that in particular X (‘) = X. Let 1 be such that x, # 0 and xq = 0 for all
q > 1, and denote Xti) = col(x~))~=,. Then
and
AX(i) = j$i+ 1) (i =O,l,...,n -1)
x(i) = 0 P (i=O,l,..., n-l, p>l+i).
Proof. We prove the assertions by induction on i. The case of i = 0 follows from the assumption on X. Next, assume that the lemma is valid for i < n - 1, and prove it for i + 1. Note that (2.2) implies
x(i+ 1) = (s _ x@l)As) x(i), (2.3)
and therefore,
x(i+ 1) P
= p-(p)x(i+l) = x;‘l _ x j7(l)Asx(i), P
But if p > 1 + i + 1, then in particular xp = 0, and also x:1 1 = 0 (as p - 1 > 1 + i). Therefore,
x(i+l) = P 0.
Moreover, it follows from (2.1) and (2.3) that
fl(i+ 1) = (AS _ E”‘F”‘AS) x(i),
But
SA _ AS = SAE’“‘F’“’ - E”‘F’l’AS
INVERSION OF TOEPLITZ MATRICES 149
for any n-by-n Toeplitz matrix A, so
MC’+‘) = (SA - SAE’q7’“‘) x(i).
Note that F(“)X(i) = x’,” = 0, as we assume that i < n - 1. Furthermore, SAX(i) = s@i) = E(i+ 1) n
Proof of Theorem 1.4. It is clear that if the Toeplitz matrix A is invertible, then both systems of equations are solvable. Conversely, assume that both conditions are fulfilled. In particular, (2.1) holds for the column vector X = col( xr>,“= i. Thus, it follows from Lemma 2.1 that
Ay = E’“+ 1-l)
for the column vector
Y = (S - XF(')AS)"-iX. (2.4)
Furthermore, condition (b) implies the existence of a column vector Z [which is precisely col( z,>,“= rl for which
It follows from Theorem 1.3 that the matrix A is indeed invertible. The formula for A-’ is a simple consequence of the identity between (2.4) and (1.4). W
In Theorem 1.4 there are two extreme cases. The first one is when 1 = n. This is precisely the case that appeared in [2] and is stated as Theorem 1.2. The second case is when 1 = 1, which happens when the matrix is an upper triangular Toeplitz matrix. This result can be easily proved independently of our main result. However, we state it for illustrative purposes.
COROLLARY 2.2. Let A = (a,_ 9 )F, 9 = 1 be a Toeplitz matrix. There exists a complex number x such that
A . [ @)] = E(l)
if and only if A is an upper triangular matrix (i.e., ai = 0 for i > 0) and a, # 0. In this case, x = l/a,, the matrix A is invertible, and its inverse
150 GEORGE LABAHN AND TAMIR SHALOM
matrix A-’ is the upper triangular Toeplitz matrix
Yn-1 **. Yl
in which
COY Y&1 = ;
-a’, -a’, **’ -a’(._,) 0
1 0 . . . 0 0 0 1 ‘. : :
0 0
0 0 ..: 1 0
n- 1
E(l)
where a’+ = a-,/a,, for i = 1,2,. . . , n - 1.
Clearly, an analogous result holds for lower triangular Toeplitz matrices as well. In general, for regular Toeplitz matrices the last column of the inverse can replace the first one. Also in this case, we might need an additional column.
THEOREM 2.3. The Toeplitz matrix A = ( ap _ y>;, y = 1 is invertible if and only if the following conditions hold:
(a) There exists a solution for AX = EC”). (b) Denote X = col( xP>p”= 1, and let m (1 < m < n) be such that x,,, z 0
and xq = 0 for all q < m. Then there exists a solution for AZ = EC”-“‘).
In this case, denote Z = col(z,),“= 1; then the inverse is given by
1 -
x7n
x, X,-l *-* x1 . . .
0 x, *** x2 y2y’ z1 . . . I: 0
: .. :I[ : . . : 0
0 0 : (j (j ..: x; y" _‘Z”_l a**. Y2 1 Zl i, 1 Z”
0
+ I. Z,-l-Y, ... Zl--y2 0 . . . 0 o\
2, . . . 52. - Y3 x1 *** 0 0 : .. : ’ . . .
0 (j ..: s, II:. x”; 1 ..: ;, I. (j
INVERSION OF TOEPLITZ MATRICES 151
where the column Y = col( y,>,“= 1 is defined by
Y = (ST - XZ7(“)AST)‘11- lx,
Proof. The proof follows essentially from Theorem 1.4, noting that for any Toeplitz matrix A, its transposed matrix AT (which is clearly a Toeplitz matrix as well) satisfies
for the n-by-n matrix
A = ]A’]
0 0 ... 0 1
0 : .f. 1 0
J= i 0 *.* 0 t
0 1 .** ; 0
1 0 ... 0 0
(2.5)
Now, Jcol(~~)“=~ = lent to (JAJkJX)
coKx,+,_ >;+ so the equation AX = E’“’ is equiva-
AT col(x >” = JE(“) 6s J” is the identity matrix), or to
n+1-D u=l = E(l). Note that x,,, is the last nonzero entry of X if and only if x[ is the’ first nonzero entry of IX for I = n + 1 - m. Apply Theorem 2.3 to the Toeplitz matrix AT, the column JX and I = n + 1 - m. Let col( Z ,,+ l_P)~=l be th e solution of the second system of equations [part (b) of Theorem 1.41. Namely, ATJZ = E (“t2-1) AZ = E”- I), or AZ = E’“-“1’.
for 2 = col(~~~)~=~. Equivalently, W
COROLLARY 2.4. Let A = (a, _ y >g, ‘I = 1 be a Toeplitz matrix. There exists a complex number x such that
A . [ xE(“)] = E’“’
if and only if A is a lower triangular matrix (i.e., a, = 0 for i < 0) and a, # 0. In this case, x = l/a,, the matrix A is invertible, and its inverse matrix A -’ is the lower triangular Toeplitz mutn’x
y1 0 ... 0
152
in which
GEORGE LABAHN AND TAMIR SHALOM
where ai = ai/aO for i = 1,2,. . . , n - 1.
As before, the above result is stated for illustrative purposes only, since the formula is easily derived independently of Theorem 2.3.
3. INVERSE FORMULAE FOR HANKEL MATRICES
Let A be an n-by-n Hankel matrix
a1 a2 a3 *** a”
a2 a3 a4 : a,+,
A= a3 : *a* .-* ani
an :
a, ant1 an+2 -*a fJ2n-1
with a,, . . . , a2n_1 complex numbers. We use the shorthand A =
(U p + 4 _ l);, 4 = 1 to denote these matrices. Hankel matrices have many features in common with Toeplitz matrices. In this section we state and prove the corresponding versions of Theorem 1.4 and 2.3 along with their corollaries as they apply to Hankel matrices.
THEOREM 3.1. and
The Hankel matrix A = (u~+~_ 1>,“, 4= 1 is invertible if
only if the following conditions hold:
(a) There exists a solution for the following system of equations:
2 ap+g-lx”-q+1= p.1 6 (p = 1,2 ,..., n). q=1
INVERSION OF TOEPLITZ MATRICES 153
(b) Let 1 be such that x1 z 0 and xy = 0 for all q > 1. Then there exists a
COHOLLARY 3.4. Let A = (a,,,, _ ,>;. ‘,= 1 be a Hankel matrix. There
exists a complex number x such that
A . [ &‘)I = E’“’
if and only if A is a lower quasitriangular matrix (i.e., ai = 0 for i < n) and
a,, # 0. In this case, x = l/a,, the matrix A is invertible and its inverse . matrix A-’ is the lower quasitriangular Hankel matrix
A-1 = : : ” yz y1 *--
Yl 0 . . .
in which
0 --a;,_, ... -a’ n+2 -a’ II+1
0 0 . . . 0 1
1 0
0 0 *-- ; ;
0 1 . . . 0 0
1- 1
E(l)
where aI = a,/a,, for i = n + 1, n + 2,. . . ,2n - 1.
Proof. Theorems 3.1 and 3.3 along with their corollaries are all proved in a similar fashion.
For each i, j = 1, . . . , n, define hi = a,_,. if / is the n-by-n matrix defined by (2.51, then
AJ = (n^,-,);,q_l = i
156 GEORGE LABAHN AND TAMIR SHALOM
a Toeplitz matrix. Parts (a) and (b) of Th eorem 3.1 then correspond to parts A (a> and (b) of Theorem 1.4 applied to the matrix A. Since J is its own inverse, Theorem 3.1 follows directly from Theorem 1.4 along the observa- tion that
A-’ = (&-I +-I,
Similarly, parts (a) and (b) of Th eorem 3.3 correspond to parts (a) and (b) of Theorem 2.3 applied to A. 4
4. MINIMAL NUMBER OF STANDARD EQUATIONS THAT DETERMINE INVERTIBILITY
In this section, we show that the results stated above are minimal. Namely, one cannot determine the invertibility of a Toeplitz or a Hankel matrix through the solvability of precisely on= system of equations of the form
AX = E,
where E is a unit column vector, unless A is triangular (Toeplitz) or quasitriangular (Hankel).
It has been shown in [5] that one cannot determine the invertibility of an n-by-n Toeplitz matrix A from the existence of solutions to AX = U, for any n - 1 fixed vectors Vi, U,, , U,, _ I E C”. Here we consider a slightly different problem. We are given two unit columns (e.g. E(l) and E’” + 2-1) as in Theorem 1.4), and we have shown that the invertibility of A can be deduced, using the structure of the solution columns (in fact, we merely used the fact that xI was the last nonzero entry of the column X which solves AX = E”‘).
THEOREM 4.1. Let X = col(x >“= 1 be a nonzero column. Let 1 and m
be two integers (1 < 1, m Q n> sue fz ihat x, # 0, x,,, # 0, and x = 0 when-
everq <lorq >m. LetE ci) be a unit column (1 < i < n). The: there exists a singular n-by-n Toeplitz matrix A = (a _ >p”, v= 1 such that AX = EC’)
unlessi=l=m=I ori=l=m=n,w zc occursifandonlyifAisa K. K.
regular traigular Toeplitz matrix (upper in the first case and lower in the
second case).
Proof. First let us consider the case in which 1 < i. Consider the Toeplitz matrix A = cap _ 4 I;, 4 = I in which aj = 0 whenever j < i - 1, a, _,
INVERSION OF TOEPLITZ MATRICES 157
= l/x,, and all the other entries are defined recursively by
aj = -cpr: a_kS,+k
s, (j > i - 1).
This Toeplitz matrix is clearly singular (having zero row as its upper row), and
it satisfies AX = E”‘.
The second case is of i < m. Now we consider the Toeplitz matrix
A = (+,);:,<,=, in which (I, = O whenever j > i - m, ni ,I, = l/s ,,,, and
all the other entries are defined recursively by
n, = -C:‘i:a.,+g,,,-k
s (j < i - tn).
!I,
This Toeplitz matrix is again a singular matrix (having zero row as its lower
row) which satisfies AX = E”‘. The last possible case is the one in which 111 < i < 1. Hut it is de;]; ;il;‘t
1 < m, so in this case X = .rE (i) for some nonzero complex number .s-. 1 I 1 < i < n, then one can take the Toeplitz matrix A = (a,’ ,, 1;:. ,, = , in which
= a,, = n ;i: ’ . .
_(,, ,) = l/x and (I, = 0 whenever j # n - 1, 0, - ( )I - 1).
1s mgular (the first and last rows are identicalj Toeplitz matrix clearly
satisfies A . [SE(‘)] = EC’).
Note that if i = 1 or i = n, then the matrix A must be an invertible
triangular Toeplitz as discussed in Corollary 2.2 and in Theorem 2.3. n
For Hankel matrices we have an analogous result, which we state without
proof.
THEOREM 4.2. Let X = col(r,,_,,+ ,I;:=, he 0 nonzero colun~n. Let 1 ad
nl he two integer (1 < 1, 1)~ < 11) mch that x, # 0, x,,, f 0, and s,, = 0 wheneoer y < 1 or q > tn. Let E (I) he n unit column (1 < i < n>. Then there
e1i.st.s n singular n-by-n Hankel nutria A = ((I,,, , _ ,>;:, ,, = , mch tht AX =
EC” unless i = 1 = tn = 1 or i = 1 = tn = n, which OCCI~TS if and ml!/ if A is
a regular cpnsitriangrtlar Hnnkel wltn’s (upper in the fir.rt cnse nnrl lower in the .second case>.
REFERENCES
1 A. Ben-A&i and T. Shaloln, On inversion of Toeplitz and close to Toeplitz matrices, LIIZWIT Algehrcr Appl. 7.5: 173-192 (1986).
2 I. C. Gohherg and N. Ya. Krupnik, A fornwla for the inversion of finite Toeplitz
matrices (in Russian), Mat. lsslcrl. 7(2):272-283 (1972).
158 GEORGE LABAHN AND TAMIR SHALOM
3 I. C. Gohberg and A. A. Semencul, On the inversion of finite Toeplitz matrices and their continuous analogs (in Russian), Mat. Issled. 7(2):201-223 (1972).
4 G. Heinig and K. Rost, Algebraic methods for Toeplitz-like matrices and opera-
tors, in Operator Theorrj, Vol. 13, Birkhauser, 1984. 5 C. R. Johnson and T. Shalom, Information requirements for Toeplitz inversion,
Linear and Multilinear Algebra, to appear. 6 G. Labahn, D. K. Choi, and S. Cabay, Inverses of block Hankel and block Toeplitz
matrices, SIAM J. Comput. 19(1):98-123 (1990). 7 L. Lerer and M. Tismenetsky, Generalized Bezoutians and the inversion problem
for block matrices, Integral Equations Operator Theory 9:790-819 (1986). 8 A. L. Sahnovic, On a method of inversion of finite Toeplitz matrices (in Russian),
Mat. Issbd. 8(4):180-186 (1973).
Received 13 ]anua y 1992; jhal manuscript accepted 29 September 1992
